Beyond scalar losses: calibrating segmentation models via gradient vector field surgery

Medical image segmentation often relies on region-based loss functions like the Dice loss to handle extreme class imbalance. However, these losses are known to produce overconfident, miscalibrated probability maps, which limits their reliability in high-stakes clinical decision-making, such as defining tumor resection margins. The authors investigate why these losses fail to produce well-calibrated probabilities and propose a method to fix this without abandoning the benefits of region-based optimization.

The authors analyze the gradient dynamics of region-based losses and demonstrate that they push predicted probabilities toward extreme values (0 or 1) regardless of the actual uncertainty. They identify that the partial derivatives of these losses w.r.t. voxel logits do not account for the prediction error magnitude, unlike the Cross-Entropy loss. To address this, they introduce a 'gradient surgery'—a modification to the gradient vector field that scales the partial derivatives linearly with the prediction error $|y - p|$. This intervention is designed to be compatible with any region-based loss function, effectively forcing the model to learn better-calibrated probabilities while maintaining the overlap-optimization strengths of the original loss.

Empirical evaluations across 2D and 3D medical imaging datasets (including breast masses, vessel segmentation, and brain metastases) show that the proposed surgery consistently improves calibration metrics—such as Expected Calibration Error (ECE) and Negative Log-Likelihood (NLL)—compared to standard Dice, Tversky, and compound losses. Crucially, this calibration improvement is achieved without degrading the Dice Similarity Coefficient (DSC), indicating that the model retains its segmentation performance while becoming more reliable in its confidence estimates.

This work provides a theoretically grounded and practically simple way to improve the reliability of deep learning models in clinical workflows. By enabling better-calibrated segmentation, the method allows clinicians to better interpret model uncertainty, which is vital for tasks where the model's output serves as a basis for surgical planning or radiotherapy targeting.

AI-generated third-party summary by ResearchPod. Not official content or an endorsement by the paper authors or affiliated organizations.